Prompt

The system prompt is the persona and the rules. The language model reads it on every single turn of a call, so it is the largest lever you have over how the agent behaves and how the call feels. You edit it on the Prompt page; it autosaves as you type.

What belongs in the prompt

The prompt describes who the agent is and how it should behave — not what it can do. What it can do lives in tools. Keep the two separate:

• Persona — name, role, tone, and how it should sound to a caller.
• Rules — what it must always do, what it must never do, how to handle the awkward cases (caller is angry, asks for something out of scope, gives a detail that does not match).
• Boundaries — when to transfer to a human, when to end the call, what it should refuse.

Prompt before, tools after

When you want the agent to behave differently, reach for the tool description first and the prompt second. A tool's description is read by the model the same way the prompt is, and it sits right next to the action — so tightening a tool description is usually a sharper fix than adding another paragraph to the prompt. Use the prompt for things that span the whole call.

Keep it tight

Long prompts are not more reliable prompts. Every rule competes for the model's attention, and contradictory or redundant instructions make behaviour less predictable, not more. Prefer:

• Short, concrete rules over long hedged ones.
• Examples of the behaviour you want over abstract descriptions of it.
• One clear instruction over three that nearly say the same thing.

Iterating

Change the prompt, then open the test panel and talk to the agent. Watch for the failure you were trying to fix and for anything you broke. When a change holds up across several runs — not one lucky pass — keep it. For graded, repeatable checks across many scenarios, see Evals.

Edits land in a draft. They only reach live callers when you publish.